Research map / Point clouds / Point cloud classification
FH-Net: A Fast Hierarchical Network for Scene Flow Estimation on Real-World Point Clouds
Lihe Ding, Shaocong Dong, Tingfa Xu, Xinli Xu, Jie Wang, Jianan Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Trans-up layer (with feature interaction) in place of linear interpolation upsampling (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.459 -> 0.255 | Improved |
| add feature interaction to the Trans-up layer (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.314 -> 0.255 | Improved |
| add hierarchical loss on intermediate flows (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.306 -> 0.265 | Improved |
| raise foreground weight to 5 and background to 0.5 in hierarchical loss (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.265 -> 0.255 | Improved |
| use Trans-flow layer instead of averaging candidate flows (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 1.384 -> 0.255 | Improved |
| use Trans-flow layer instead of nearest point offset (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 1.025 -> 0.255 | Improved |
| add position encoding in Trans-flow layer (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.278 -> 0.255 | Improved |
| add residual flow refinement in flow propagation (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: 0.255 -> 0.223 | Improved |
| add copy-and-paste augmentation with 3 cars, 1-5 pedestrians, 1-4 cyclists (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI), overall EPE3D | EPE3D: 0.308 -> 0.255 | Improved |
| add copy-and-paste augmentation with 3 cars, 1-5 pedestrians, 1-4 cyclists (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI), foreground | foreground EPE3D: 0.278 -> 0.127 | Improved |
| increase pasted cars from 3 to 6 (ablation) | FH-L · Lidar-KITTI (trained on SF-KITTI) | EPE3D: foreground 0.127 -> 0.106; overall 0.255 -> 0.259 | Mixed |
| train FH-L on SF-Waymo instead of FT3D (ablation) | FH-L · SF-Waymo/eval | EPE3D: 0.673 -> 0.243 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- FlowNet3D — FlowNet3D: Learning Scene Flow in 3D Point Clouds (CVPR 2019) · better on 5 of 5 shared results
- PointPWC-Net — PointPWC-Net: Cost Volume on Point Clouds for (Self-)Supervised Scene Flow Estimation (ECCV 2020) · better on 3 of 4 shared results
- FESTA — FESTA: Flow Estimation via Spatial-Temporal Attention for Scene Point Clouds (CVPR 2021) · better on 1 of 1 shared result
- HPLFlowNet — HPLFlowNet: Hierarchical Permutohedral Lattice FlowNet for Scene Flow Estimation on Large-Scale Point Clouds (CVPR 2019) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PointFlowNet: Learning Representations for Rigid Motion Estimation From Point Clouds (CVPR 2019 · similarity 0.87)
- HPLFlowNet: Hierarchical Permutohedral Lattice FlowNet for Scene Flow Estimation on Large-Scale Point Clouds (CVPR 2019 · similarity 0.87)
- IHNet: Iterative Hierarchical Network Guided by High-Resolution Estimated Information for Scene Flow Estimation (ICCV 2023 · similarity 0.86)
- SCTN: Sparse Convolution-Transformer Network for Scene Flow Estimation (AAAI 2022 · similarity 0.86)
- FlowMamba: Learning Point Cloud Scene Flow with Global Motion Propagation (AAAI 2025 · similarity 0.84)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
